Cross-batch Reference Learning for Deep Classification and Retrieval

by Huei-Fang Yang, Kevin Lin, and Chu-Song Chen

Introduction

This code implements the cross-batch reference (CBR) learning as described in our ACM MM 2016 paper.

Citing the CBR

If you find CBR useful in your research, please consider citing:

@inproceedings{yang:acmmm16,
    author    = {Huei-Fang Yang and
                 Kevin Lin and
                 Chu-Song Chen},
    title     = {Cross-batch Reference Learning for Deep Classification and Retrieval},
    booktitle = {Proc. ACM MM},
    pages     = {1237--1246},
    year      = {2016}
}

Requirements

1. Caffe, matcaffe and pycaffe (see: Caffe installation instructions)
2. MATLAB (required for performance evaluation)

Train a network with CBR on CIFAR-10

1. Download the pretrained model and CIFAR-10 dataset.

   $./download.sh
   

2. Modify the CAFFE_BIN in run.sh to the path where the CAFFE is installed.

3. Modify the caffe_root in update_model.py to the path where the CAFFE is installed.

4. Launch the script to train a network with CBR. This would take a few hours.

   $./run.sh
   

Evaluate the trained model

1. Modify the addpath in run_cifar10.m to the path where the CAFFE is installed.

2. Launch MATLAB and run the evaluation code to obtain mAP and precition@k.

   >>run_cifar10
   

Contact

Please feel free to contact Huei-Fang Yang (hfyang@citi.sinica.edu.tw), Kevin Lin (kevinlin311.tw@iis.sinica.edu.tw), or Chu-Song Chen (song@iis.sinica.edu.tw) if you had any questions.